I lately peruse End Times Fascism and the Fight for the Living World by Naomi Klein and Astra Taylor. It's an outstanding publish and I extremely propose that everyone peruse it. Tech oligarchs create up an crucial component of the end-times-fascist coalition. One item that stood out to me, as a application and safety engineer who finds LLMs to be extremely powerful, is that the American frontier labs are run by unhinged lunatics who genuinely accept they're creating the Machine God, that they're fine alongside burning downward the earth to bring concerning AGI, and that they accept alongside religious fervor that they're the lone ones astute adequate to safely authority it. I already benevolent of knew this, but this publish really made it descend in.
As much as the AI hype all about us pretends that OpenAI and Anthropic are distinctive and special, they're not. They're not the lone companies doing this work. In fact, a lot of the open importance models are fair concerning as fine as the latest ChatGPT and Claude models, including for incredibly complex tasks akin exposure research.
Unlike models from OpenAI and Anthropic, open importance models have important advantages:
- It's imaginable to run them privately on your own hardware (if you can somehow get a device alongside adequate RAM)
- You can use a provider akin Tinfoil that relies on safe enclaves to run them privately (I conversation concerning this in a bit additional item below)
- If privacy isn't a huge concern, you can use any figure of another providers, several of them fairly affordable, and it's uncomplicated to toggle so you're not locked into a sole evil vendor
- You can motionless use mighty LLMs without financially supporting deranged oligarchs who wager alongside existence on Earth so that they can continually remain in power
ChatGPT and Claude volition never be private
These companies lone recommendation closed versions of their mighty models (despite the term "open" in "OpenAI"). You pay them for conversation and/or API access, and they are allowed to appearance at everything you do.
They recommendation a checkbox so you can opt out of them using your data for training, but your conversation sessions motionless aren't private. They have admission to it all. They can appearance at it, portion it alongside governments or companies, or do any they want. It's possible, apt even, that they won't use it for training if you've opted out, at smallest not correct now. But do you really rely that long-term? Have they done item to create anyone accept that they regard people's data?
OpenAI and Anthropic are company gatekeepers
Their models have strict guardrails, as anyone who's tried to use them for item cybersecurity-related has no doubt noticed.
Every meeting has a cop delegate study your copy complete your shoulder, making certain you're not using it for hacking. If it thinks you power be, the cop delegate stops your meeting from continuing. It's extremely strict, and false positives happen all the time. Of course, not everyone has the cop observing them.
Some group get admission to the models alongside looser restrictions or alongside none at all. OpenAI has Trusted Access for Cyber and Anthropic has the Cyber Verification Program. To acquire admission to additional capable, small policed versions of their models, individuals and companies need to use and be approved by OpenAI and Anthropic. A lot of US cybersecurity companies get access. Almost certainly US armed forces contractors do too, and likely any parts of the US authorities that desire it.
While there's several deserve in restricting mighty models, handing the supreme power to decide who gets admission to them to Sam Altman and Dario Amodei – alongside consultation from the Trump Administration, of way – is a terrible, undemocratic idea.
I think a improved method is to merely not put yourself in a circumstance anywhere you depend on ChatGPT or Claude. Instead of begging company overlords for privileged access, it's improved to expend your period figuring out how to get models akin GLM-5.3 to encounter your needs. And it's not hard. These models are good, alongside the added advantage that they don't have a cop observing complete your shoulder!
Open importance models are impressive
So far, open importance models appear to be concerning four months rearward the fancy proprietary models from OpenAI and Anthropic.
Four months is nothing.
I'm extremely impressed correct now alongside GLM-5.3. There's plentifulness of data and benchmarks out there, but my vibes-based opinion from using it is that it's basically on par alongside Claude Opus 5.5 for application engineering.
I'm additionally impressed alongside GLM-5.3-Flash. This is akin to Claude Opus 4.8. It's much smaller and cheaper than GLM-5.3, and motionless incredibly capable. It's additionally really feasible for an idiosyncratic individual to get hardware to run it, akin if you can one day get your hands on a Mac Studio alongside 512GB of RAM.
GLM models are made by a Chinese business called Z.ai. While the models are impressive, I don't propose making an document immediately alongside Z.ai or sending them any of your data. Their coding harness ZCode (their type of Claude Code) was recently caught silently uploading local origin code repos and sending them rear to Alibaba Cloud without consent. Also, it's harmless to assume that Chinese companies volition portion any data you provision to them alongside the Chinese government.
Other open importance models from Chinese companies that are notable contain the DeepSeek, Qwen, and Kimi families of models. It's absolutely fine, safe, and safe to use these open importance models. Just don't rely the companies themselves alongside your conversation sessions.
Another frontier AI lab value a appearance is Mistral. As a French company, it's the lone big participant that's not Chinese or American. They fair released an open importance frontier example called Mistral Large 4, aka Le Chonk. I think it's fairly a bit additional sensible to rely Mistral alongside your data compared to the Chinese or American companies.
Private conclusion alongside Tinfoil or Maple AI
It's imaginable to have personal AI, equal without hosting it locally yourself, using safe enclaves. These aren't perfect, but I do think they provision a extremely meaningful difference in privacy compared to the position quo mode, which is "the provider can admission everything alongside no restrictions."
The business that's doing the most activity current is Tinfoil. If you desire to study the details, peruse their docs on safe enclaves and distant attestation. Another business doing this is Maple AI – although, technically, Maple uses Tinfoil for conclusion in the background.
Both Tinfoil and Maple are fine options. Tinfoil is cheaper, but Maple is additional consumer-focused.
If you desire a ChatGPT-like experience, anywhere you can conversation to a chatbot, have it appearance material up for you on the internet, provision it admission to docs to peruse or do material with, admission it in a mobile app, etc., afterward I'd go alongside Maple. There's a small liberated scheme and a Pro scheme that expenses $20/month.
If you desire to do genuine agentic coding tasks akin application engineering, reverse engineering, malware analysis, CTFs or pen tests, etc., afterward you should use Tinfoil instead. They accusation by the token for personal inference, so no monthly plan. GLM-5.3 currently expenses $1.80 per 1M input tokens, $0.80 per 1M cached tokens, and $5.75 per 1M output tokens. You can use considerably small tokens if you choose the "low" reasoning level, and it's motionless extremely impressive.
Inference without the privacy
If you don't desire to pay a premium for personal conclusion from Tinfoil or Maple, but you motionless desire to liberated yourself from OpenAI and Anthropic, there are lots of less-private options that are additionally extremely cheap.
One that I've been playing alongside is OpenCode Go. It expenses $10/month for additional than $10 value of credits. This is an conclusion router, which basically method your API queries do item akin this: your device -> OpenCode Go API -> the genuine conclusion provider. This method that the two OpenCode Go and any backend provider they're using volition have admission to your meeting data. That said, the privacy section of their docs promises which models data won't be used for training (most of them) and what the data preservation is (zero days for most, including all the GLM, Qwen, Kimi, and DeepSeek models). Of course, it's fair a promise. Nothing specialized guarantees it.
There are plentifulness of another cheap options out there for non-private inference, if you fair desire to test alongside mighty open importance models.
How I'm using these models
I've been using the two Tinfoil and OpenCode Go, configured alongside my current harness of choice, OpenCode. I've been mostly using GLM-5.3 on low reasoning, alongside awesome results. (Things alter so quickly alongside the open importance models, who knows what volition be finest six months from now.)
I use OpenCode Go conclusion for entirely community work, akin whenever I'm implementing item that's already open source.
I use Tinfoil for everything else.
I've lone fair started operating this way, but so far, I think the open importance models can effortlessly entirely substitute my use of ChatGPT and Claude, at a comparable or cheaper price, alongside the added advantage that they're far additional personal and there are far small guardrails.
